|Year : 2021 | Volume
| Issue : 1 | Page : 57-63
Machine Learning Approaches for Prognostication of Newly Diagnosed Glioblastoma
Thara Tunthanathip, Thakul Oearsakul
Division of Neurosurgery, Department of Surgery, Faculty of Medicine, Prince of Songkla University, Hat Yai, Songkhla, Thailand
|Date of Submission||22-Sep-2020|
|Date of Decision||21-Oct-2020|
|Date of Acceptance||22-Nov-2020|
|Date of Web Publication||12-Feb-2021|
Division of Neurosurgery, Department of Surgery, Faculty of Medicine, Prince of Songkla University, Hat Yai, Songkhla 90110
Source of Support: None, Conflict of Interest: None
| Abstract|| |
Background: Glioblastoma (GBM) is the poorest prognosis in glioma. Although Temozolomide (TMZ) with radiotherapy following tumor resection is currently the standard treatment, the high cost has become an economic burden in a limited-resource setting. In an era of disruptive innovation, machine learning (ML) has been recently performed to be the clinical prediction tool for prognostication, especially GBM. The aim of the study was to assess the predictability of ML algorithms for 2-year survival in patients with GBM. Methods: A retrospective cohort study was performed in patients with GBM. Various clinical, radiological, and treatment variables were collected, and the outcome was a 2-year living status as bi-classifiers. The candidate variables, which had a P<0.1, were performed to train the ML model. For training the ML model, random forest (RF), logistic regression (LR), and support vector machines were used for training the model and testing the predictive performance. Results: There were 190 GBM patients in the cohort. Four candidate variables were used for building the ML model and testing the performance of each algorithm. The LR and RF algorithms had an acceptable performance for predicting a 2-year survival with an area under the receiver operating characteristic curve at 0.82 and 0.81, respectively. Conclusion: ML-based algorithms had an acceptable performance for the prognostication of 2-year survival in GBM patients that could be implicated in real-world practice for selecting patients with a favorable prognosis and developing treatment strategies for resource allocation in a limited-resource setting.
Keywords: Disruptive innovation, glioblastoma, machine learning, random forest, support vector machine
|How to cite this article:|
Tunthanathip T, Oearsakul T. Machine Learning Approaches for Prognostication of Newly Diagnosed Glioblastoma. Int J Nutr Pharmacol Neurol Dis 2021;11:57-63
|How to cite this URL:|
Tunthanathip T, Oearsakul T. Machine Learning Approaches for Prognostication of Newly Diagnosed Glioblastoma. Int J Nutr Pharmacol Neurol Dis [serial online] 2021 [cited 2021 Mar 7];11:57-63. Available from: https://www.ijnpnd.com/text.asp?2021/11/1/57/309285
| Introduction|| |
Glioblastoma (GBM) is the most common malignant brain tumor in adults and the poorest prognosis among gliomas, and a 2-year survival probability has been mentioned for this tumor. According to Stupp et al., patients with GBM who received concurrent chemoradiotherapy had a 2-year survival probability of 26.5%, while patients who received radiotherapy alone had a 2-year survival rate of 10.4%. Furthermore, the median survival time of GBM has been reported in 11 to 15 months following tumor resection with adjuvant chemoradiotherapy.,,
Currently, although targeted therapy has been investigated for improving the survival time, real-world situations have the limitation of applying a high-cost treatment., Thus, novel treatment has become an economic burden in middle-low income countries.,, Prognostication of disease is also one of the essential processes for resource allocation—treatment strategies based on a favorable prognosis of a disease challenge to manage this economic burden., Therefore, various factors have been studied and reported that were associated with the prognosis in prior studies; such as the Karnofsky performance status (KPS),, extent of resection, and concurrent chemoradiotherapy with Temozolomide (TMZ).
Machine learning (ML) has become one of the valuable clinical prediction tools in an era of disruptive innovation. From the literature review, previous clinical studies used various ML algorithms for classifying and predicting the clinical outcome and prognostication of various diseases; such as neuro-oncology,, traumatic brain injury, stroke, and postoperative complications. Deep learning and convolutional neural networks have been used in glioma to classify the tumor grade and molecular subtype based on image processing and computer vision.,, The ML model using logistic regression (LR) had a sensitivity of 0.97, accuracy of 0.93, and area under the receiver operating characteristic curve (AUC) of 0.94 for classifying between low-grade glioma (LGG) and GBM, whereas Wu et al. used the ML model by using a support vector machine (SVM) to classify the LGG from high-grade glioma (HGG) with an AUC of 0.987, and an anaplastic astrocytoma from GBM with an AUC of 0.992. Additionally, image-based ML algorithms; such as a convolutional neural network (CNN) have been performed to classify various grading of gliomas. In detail, Yang et al. used a CNN algorithm to distinguish LGG and HGG from magnetic resonance imaging (MRI); thus, the accuracy and AUC were 0.909 and 0.939, respectively. Furthermore, Zhuge et al. reported that a CNN algorithm was the effective method for tumor grading in gliomas with high accuracy of 0.963 to 0.971.
A few studies mentioned building the ML model for a clinical outcome; such as mortality and the overall survival of glioma or GBM from the literature review. In detail, Tan et al. used the ML techniques for the overall survival of 112 patients with HGG by combining clinical, genetic, and MRI radiomic features; therefore, the AUC of the test data set was 0.758. Moreover, Lao et al. used a deep learning-based model for predicting survival in GBM and reported an AUC of 0.710 (95% confidence interval (CI) 0.588–0.932). In the face of this gap, the authors aimed to assess the predictability of ML algorithms for the two-year survival in patients with GBM.
| Methods|| |
Study designs and study population
All patients newly diagnosed with GBM were admitted into a tertiary center of southern Thailand between January 2000 and December 2019. The exclusion criteria were patients who had unavailable histological slides for confirming the diagnosis, unavailable complete medical records, and could not assess an updated outcome. Several clinical features, radiological findings, and treatments were reviewed from computer-based medical records.
The operational definition was prepared before the medical record review. The extent of resection (EOR) was defined based on the study of Bloch et al. A residual tumor was evaluated by postoperative T1W with contrast imaging. Total resection referred to a residual tumor <5% following the operation, whereas partial resection was defined as a residual tumor of ≥5%. Moreover, a biopsy referred to the diagnostic procedure, in which there was no attempt made to remove the tumor. According to Lacroix et al., the eloquent areas were the motor cortex, sensory cortex, visual center, speech center, thalamus, hypothalamus, basal ganglion, dentate nucleus, and brainstem. The degree of tumor necrosis, mass effect, and enhancement were graded by prior study, and tumor volume estimations were performed according to the study of Tunthanathip et al.
The present study’s outcome was categorized as death or living status, which was confirmed from the local civil registration database on August 2, 2020. The present study was approved by the Research Ethics Committee, Faculty of Medicine, Prince of Songkla University (REC.63-372-10-1). Although the informed consent was not performed from the retrospective study design, the patients’ identification numbers were encoded before the analysis.
Patients’ characteristics, radiologic findings, and treatment were presented for descriptive purposes with proportion. The mean with standard deviation (SD) was used when they were continuous variables. For the independent variable, the updated patients’ status was categorized as binary classifiers. The difference in each variable’s distribution across the outcome groups was estimated using a chi-square test or Fisher exact test. P values < 0.05 were regarded as statistically significant. However, candidate variables that had a P < 0.10 were constructed in the ML model. In this part, the statistical analysis was performed using the R version 3.6.2 software (R Foundation, Vienna, Austria).
Seventy percent of the whole data was used to train the ML model, whereas the remaining 30% was achieved for testing the model’s predictability. Because patients who had inadequate information were excluded, missing data management did not need data processing.
Supervised ML algorithms were used for training the model with 10-fold cross-validation as follows: random forest (RF), LR, and SVM with the radial basis function (RBF) kernel, and SVM with a linear kernel. Therefore, a confusion matrix was obtained to present the performance of the model by the test data set for which the actual values of the outcome were known. In detail, the performances of each algorithm were estimated as follows: sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy. Moreover, the authors used the receiver operating characteristic (ROC) curve and AUC to estimate the discrimination of the model. In general, acceptable discrimination would be an AUC of ≥0.7, whereas good and excellent discrimination would be an AUC of ≥0.8 and ≥0.9, respectively.
| Results|| |
Clinical and radiological characteristics
Initially, a total of 207 patients were newly diagnosed with GBM. Ten patients had an unavailable histopathological-confirmed diagnosis, five patients had incomplete information from their medical records, and two patients had unavailable preoperative radiological imaging. Therefore, 190 patients were included for the analysis in the present study. The mean age was 51.2 years (SD 15.8), and 56.3% were male. In addition, more than half of the patients had preoperative KPS ≤ 70.
From the imaging features, the temporal and frontal lobes were the common location of GBM involvement with 35.3% and 28.4%, respectively, whereas 58.9% of cases involved an eloquent area. Moreover, 80.5% of GBMs were solitary, and 41.6% of the subjects had a preoperative tumor volume of >50 ml. Following the operation, total resection was performed in 20.5% patients, whereas 64.2% patients had a partial resection. Therefore, TMZ with radiotherapy following surgery was achieved in 32.1% patients, and 39.5% patients had postoperative KPS ≥ 80. In addition, [Table 1] shows the baseline characteristics according to the survival status.
|Table 1 Baseline characteristics of 2-year survivors and nonsurvivors (N = 190)|
Click here to view
Clinical characteristics and imaging features, which had a P value of the chi-square and Fisher exact tests of ≤0.1, were used for the supervised ML. However, the temporal and frontal locations were parts of the eloquent area category. These variables were removed before the development of the ML model. Therefore, the eloquent area, the extent of resection, adjuvant therapy, and postoperative KPS were candidate variables for building the ML model.
After splitting the data, 133 patients were included for building the ML model, and the remaining 57 patients were tested for the model performance. Because preoperative KPS was only a variable with a P > 0.05, a sensitivity analysis was conducted to compare the model’s performances with/without this variable. The results of the performance are shown in [Table 2].
|Table 2 Comparison performance of each algorithm between the first model and the second model for 2-year survivors.|
Click here to view
Overall, the ML model with preoperative KPS had a better performance for predicting the 2-year survivors than all algorithms of the ML models that did not include preoperative KPS [Figure 1]. For the model with postoperative KPS, the LR algorithms with the regularization parameter C at 0.01 had a sensitivity of 1.00, PPV of 0.89, and accuracy of 0.89. RF with 100 trees in the forest and five depths of tree, sensitivity, PPV, and accuracy were 0.94, 0.88, and 0.84, respectively. Although sensitivity, PPV, and the accuracy of the RBF kernel SVM algorithm with the regularization parameter C at 0.01 had similar performances with the RF, the AUC of the RF and RBF kernel SVM was 0.81 and 0.58, respectively.
|Figure 1 Receiver operating characteristic curve and area under the curve of each machine learning algorithm. (A) Random forest with the first model. (B) Random forest with the second model. (C) Logistic regression with the first model. (D) Logistic regression with the second model. (E) Radial basis function kernel support vector machine with the first model. (F) Radial basis function kernel support vector machine with the second model. AUC, area under the receiver operating characteristic curve; SVC, support vector machine.|
Click here to view
| Discussion|| |
Factors associated with the prognosis in the present study were the eloquent area, the extent of resection, and TMZ; additionally, the authors’ results were concordant with earlier studies.,,,,,, The EOR was reported as the prognostic factor in the literature when the volumetric EOR reached 95% to 98% of the tumor resection. However, not all patients were amenable to total resection that ranged from 22% to 60% according to prior studies., Tumor location is one of the indicators that neurosurgeons consider selecting an operation. An inaccessible tumor location would lead to limited EOR that would comprise bilateral GBM, deep-seated tumors, and eloquent areas., Therefore, intraoperative brain mapping techniques were performed for maximizing resection, whereas minimizing morbidity. In addition, TMZ with radiotherapy significantly increases the survival of GBM patients following surgery. Rønning et al. conducted a population-based study to evaluate the effect of TMZ in GBM patients and found that the median overall survival for TMZ with radiotherapy and radiotherapy alone was 16.2 and 9.0 months, respectively. Thus, TMZ with radiation had a potential benefit; this chemotherapy has become the standard treatment in several countries.,
However, the expensive cost of TMZ treatment would result in an economic burden in a limited-resource setting. This limitation would lead to knowing how it would be possible to afford effective standard treatment. Selecting patients and predicting a favorable prognosis was mentioned for the likelihood of cost-effectiveness in a limited-resource setting. Hence, prediction tools have an accuracy of predictability that were studied in the literature.
The prognostic nomogram is one of the clinical prediction tools that has been used for the pre-notification in glioma. Yang et al. used a nomogram for predicting the survival in HGG and found an AUC of 0.738, and prognostic nomogram for patients with diffuse astrocytoma had an AUC of 0.75 according to a prior study. Alternatively, ML has been used for building the predictive model in previous clinical studies. Lao et al. predicted the survival of GBM patients using a deep learning-based model that had an AUC of 0.71, whereas Peeken et al. used a combination of various clinical characteristics and MRI-based, pathological information of 189 GBM patients to build the ML model. Therefore, the performance of the ML model for overall survival prediction was an AUC of 0.73. For prognostication, most of the predictive model’s performance was accepted in the range of 0.7 to 0.8.,,,
In the present study, an AUC of 0.80 was an acceptable performance, and high levels of sensitivity and PPV could be implicated in general practice as a screening tool. Furthermore, the authors compared the results between the models with and without postoperative KPS. This factor had a P < 0.1 but P >0.05; nonetheless, this factor directly affected the performance and AUC of the predictability. Gravesteijn et al. mentioned the number of parameters that were directly associated with the model’s performance. A large number of predictors became high-dimensional data and selection of more hyperparameters was done for the model adjustment. ,,
To the best of the authors’ knowledge, this research was the first study to reveal an acceptable performance of the ML-based model for prognostication in glioma. This clinical prediction tool could be implicated in predicting the prognosis in real-world practice. However, the limitations of the present study were acknowledged. First, the authors considered a small number of patients for training and testing the ML model. Furthermore, the authors later realized that they had discarded the greater potential of the study by excluding patients who had missing parameters. Multicenter research or meta-analysis would resolve this limitation by increasing the number of the study population. Obtaining more training data would aid to improve the learning of the models, particularly supervised learning; hence, the performances of these ML models would be more effective for the prediction. Moreover, the large number of training data would result in obtaining more labeled data (outcome) that would encourage the performances of the supervised learning algorithms. Second, the study design was a retrospective approach that led to selection and information bias. However, the authors attempted to manage this limitation by setting an operational definition, inclusion, and exclusion criteria before the review. Finally, the ML model should be estimated with a greater number of unseen data to confirm generalizability that would be a challenge for conducting a prospective study in the future for testing these models.
| Conclusion|| |
The present study revealed a positive performance (AUC > 0.8) of the ML-based models for the prognostication in GBM. For the implication, these effective prediction tools would be applied in the future for selecting patients with a favorable prognosis and developing treatment strategies for resource allocation in a limited-resource setting.
The authors would like to offer their sincere gratitude to Assistant Prof. Kanet Kanjanapradit for confirming the GBM diagnosis from the histological slides.
All procedures performed in the study that involved studies requiring human participants were in accordance with the ethical standards of the institutional and/or national research committee or both and with the Declaration of Helsinki 1964 and its later amendments or comparable ethical standards.
TT and TO conceived the study and designed the method. TT supervised the conduct of the data collection. TT and TO undertook the recruitment of the participating centers and patients and managed the data, including quality control. TT provided statistical advice on the study design and analyzed the data, and drafted the manuscript. All authors contributed substantially to its revision. TT takes responsibility for the paper as a whole.
Some study population were obtained from the study by Tunthanathip et al., and GBM patients diagnosed in 2019 who died before August 2020 were added in this study. However, this study focused on a 2-year survival predictability of ML algorithms.
Financial support and sponsorship
Conflicts of interest
There are no conflicts of interest.
| References|| |
Davis ME. Glioblastoma: overview of disease and treatment. Clin J Oncol Nurs 2016;20(suppl 5):S2–S8.
Stupp R, Mason WP, van denBent MJ et al.
Radiotherapy plus concomitant and adjuvant temozolomide for glioblastoma. N Engl J Med 2005;352:987-96
Le Rhun E, Preusser M, Roth P et al.
Molecular targeted therapy of glioblastoma. Cancer Treat Rev 2019;80:1-13.
Lamers LM, Stupp R, van den Bent MJ et al.
Cost-effectiveness of temozolomide for the treatment of newly diagnosed glioblastoma multiforme: a report from the EORTC 26981/22981 NCI-C CE3 Intergroup Study. Cancer 2008;112:1337-44.
Kovic B, Xie F. Economic evaluation of bevacizumab for the first-line treatment of newly diagnosed glioblastoma multiforme. J Clin Oncol 2015;33:2296-302.
Chen Z, Zhan M, Tian F, Xu T. Cost-effectiveness analysis of the addition of bevacizumab to temozolomide therapy for the treatment of unresected glioblastoma. Oncol Lett 2020;19:424-30.
Martikainen JA, Kivioja A, Hallinen T et al.
Economic evaluation of temozolomide in the treatment of recurrent glioblastoma multiforme. Pharmacoeconomics 2005;23:803-15.
Daniels N. Resource allocation and priority setting. In Barrett DH, Ortmann LW, Dawson A et al.
, eds. Public Health Ethics: Cases Spanning the Globe [Internet]. Cham, Switzerland: Springer 2016. p. 137-76. Available from: https://www.ncbi.nlm.nih.gov/books/NBK435786/
. Accessed October 5, 2020.
Tunthanathip T, Sangkhathat S, Tanvejsilp P et al.
Economic evaluation of temozolomide for newly diagnosed glioblastoma in Thailand according to IDH1 status. Proc 11th Joint Semin Biomed Sci 2019 Nov 13-15; Krabi, Thailand. Songkhla: Prince of Songkla University; 2019.
Lamborn KR, Chang SM, Prados MD. Prognostic factors for survival of patients with glioblastoma: recursive partitioning analysis. Neuro Oncol 2004;6:227-35.
Liang J, Lv X, Lu C et al.
Prognostic factors of patients with Gliomas − an analysis on 335 patients with glioblastoma and other forms of gliomas. BMC Cancer 2020;20:35.
Stupp R, Hegi ME, Mason WP et al.
Effects of radiotherapy with concomitant and adjuvant temozolomide versus radiotherapy alone on survival in glioblastoma in a randomised phase III study: 5-year analysis of the EORTC-NCIC trial. Lancet Oncol 2009;10:459-66.
Zlochower A, Chow DS, Chang P et al.
Deep learning AI applications in the imaging of glioma. Top Magn Reson Imaging 2020;29(2);115-21.
Papp L, Pötsch N, Grahovac M et al.
Glioma survival prediction with combined analysis of in vivo 11
C-MET PET features, ex vivo features, and patient features by supervised machine learning. J Nucl Med 2018;59:892-99.
Amorim RL, Oliveira LM, Malbouisson LM et al.
Prediction of early TBI mortality using a machine learning approach in a LMIC population. Front Neurol 2020;10:1-9.
Heo J, Yoon JG, Park H et al.
Machine learning-based model for prediction of outcomes in acute stroke. Stroke 2019;50:1263-65.
Saber H, Somai M, Rajah GB et al.
Predictive analytics and machine learning in stroke and neurovascular medicine. Neurol Res 2019;41:681-690.
Wu CC, Guo WY, Chen MH et al.
Direct measurement of the signal intensity of diffusion-weighted magnetic resonance imaging for preoperative grading and treatment guidance for brain gliomas. J Chin Med Assoc 2012;75:581-8.
Li-Chun Hsieh K, Chen CY et al.
Quantitative glioma grading using transformed gray-scale invariant textures of MRI. Comput Biol Med 2017;83:102-8.
Wu J, Qian Z, Tao L et al.
Resting state fMRI feature-based cerebral glioma grading by support vector machine. Int J Comput Assist Radiol Surg 2015;10:1167-74.
Yang Y, Yan LF, Zhang X et al.
Glioma grading on conventional MR images: a deep learning study with transfer learning. Front Neurosci 2018;12:1-10.
Zhuge Y, Ning H, Mathen P et al.
Automated glioma grading on conventional MRI images using deep convolutional neural networks. Med Phys 2020;47:3044-53.
Tan Y, Mu W, Wang XC et al.
Improving survival prediction of high-grade glioma via machine learning techniques based on MRI radiomic, genetic and clinical risk factors. Eur J Radiol 2019;120:1-10.
Lao J, Chen Y, Li ZC et al.
A deep learning-based radiomics model for prediction of survival in glioblastoma multiforme. Sci Rep 2017;7:10353.
Bloch O, Han SJ, Cha S et al.
Impact of extent of resection for recurrent glioblastoma on overall survival: clinical article. J Neurosurg 2012;117:1032-8.
Lacroix M, Abi-Said D, Fourney DR et al.
A multivariate analysis of 416 patients with glioblastoma multiforme: prognosis, extent of resection, and survival. J Neurosurg 2001;95:190-8.
Tunthanathip T, Madteng S. Factors associated with the extent of resection of glioblastoma. Precis Cancer Med 2020;3:12.
Roelen CA, Bültmann U, van Rhenen W et al.
External validation of two prediction models identifying employees at risk of high sickness absence: cohort study with 1-year follow-up. BMC Public Health 2013;13:105.
Young RM, Jamshidi A, Davis G et al.
Current trends in the surgical management and treatment of adult glioblastoma. Ann Transl Med 2015;3:1-15.
Spena G, D’Agata F, Panciani PP et al.
Supratentorial gliomas in eloquent areas: which parameters can predict functional outcome and extent of resection? PLoS One 2013;8:1-8.
Rønning PA, Helseth E, Meling TR et al.
A population-based study on the effect of temozolomide in the treatment of glioblastoma multiforme. Neuro Oncol 2012;14:1178-84.
Garside R, Pitt M, Anderson R et al.
The effectiveness and cost-effectiveness of carmustine implants and temozolomide for the treatment of newly diagnosed high-grade glioma: a systematic review and economic evaluation. Health Technol Assess 2007;11:1-221.
Messali A, Hay JW, Villacorta R. The cost-effectiveness of temozolomide in the adjuvant treatment of newly diagnosed glioblastoma in the United States. Neuro Oncol 2013;15:1532-42.
Wasserfallen JB, Ostermann S, Pica A et al.
Can we afford to add chemotherapy to radiotherapy for glioblastoma multiforme? Cost-identification analysis of concomitant and adjuvant treatment with temozolomide until patient death. Cancer 2004;101:2098-105.
Wu B, Miao Y, Bai Y et al.
Subgroup economic analysis for glioblastoma in a health resource-limited setting. PLoS One 2012;7:1-9.
Yang Y, Yao M, Long S et al.
Prognostic nomograms for primary high-grade glioma patients in adult: a retrospective study based on the SEER database. Biomed Res Int 2020; 2020:1-19.
Peeken JC, Goldberg T, Pyka T et al.
Combining multimodal imaging and treatment features improves machine learning-based prognostic assessment in patients with glioblastoma multiforme. Cancer Med 2019;8:128-36.
Tunthanathip T, Ratanalert S, Sae-Heng S et al.
Prognostic factors and nomogram predicting survival in diffuse astrocytoma. J Neurosci Rural Pract 2020;11:135-43.
Peeken JC, Hesse J, Haller B et al.
Semantic imaging features predict disease progression and survival in glioblastoma multiforme patients. Strahlenther Onkol 2018;194:580-90.
Gravesteijn BY, Nieboer D, Ercole A et al.
Machine learning algorithms performed no better than regression models for prognostication in traumatic brain injury. J Clin Epidemiol 2020;122:95-107.
Beam AL, Kohane IS. Big data and machine learning in health care. JAMA 2018;319:1317-8.
Rajkomar A, Oren E, Chen K et al.
Scalable and accurate deep learning with electronic health records. NPJ Digit Med 2018;1:1-10.
Tunthanathip T, Sae-Heng S, Oearsakul T et al.
Machine learning applications for the prediction of surgical site infection in neurological operations. Neurosurg Focus 2019;47:1-9.
[Table 1], [Table 2]